## What problem are you trying to solve?

When using the `chatgpt` provider for Codex models through LiteLLM, Codex sends session metadata in the `x-codex-turn-metadata` request header.

Example header value:

```json { "session_id": "019ea92f-a0c5-7f53-8d6c-...", "thread_id": "019ea92f-a0c5-7f53-8d6c-..." } ```

However, LiteLLM does not currently promote this Codex session identifier into its standard logging/session metadata. As a result, when using `langfuse_otel`, Codex model calls are not grouped into the expected Langfuse session/trace structure unless a custom callback manually maps the header.

## Expected behavior

For requests handled by the `chatgpt` provider, LiteLLM could detect `x-codex-turn-metadata` and map:

- `session_id`, or `thread_id` as fallback

into the standard LiteLLM logging metadata, for example:

- `litellm_session_id` - `metadata.session_id` - Langfuse/OTEL session metadata - optionally trace/session id fields used by `langfuse_otel`

This would allow Codex runs to appear as coherent sessions in Langfuse without requiring a custom LiteLLM callback.

## Current workaround

We currently use a custom LiteLLM callback:

```python import json

from lit…